On Wed, 4 Feb 2009, Greg KH wrote:
> On Wed, Feb 04, 2009 at 09:32:51PM +0100, Geert Uytterhoeven wrote:
> > On Wed, 4 Feb 2009, Geert Uytterhoeven wrote:
> > > On Mon, 19 Jan 2009, Shane Hathaway wrote:
> > > > Daolong Wang wrote:
> > > > > On Mon, Jan 19, 2009 at 11:21 PM, Jeff Dike <jdike@...> wrote:
> > > > >> On Sun, Jan 18, 2009 at 02:23:46PM +0800, Daolong Wang wrote:
> > > > >>> I can confirm this link error.
> > > > >> In what environment? I see no problems here.
> > > >
> > > > I can also confirm this link error. The problem occurs when compiling
> > > > either 2.6.28.1 or 2.6.27.12; I didn't try anything earlier. The patch
> > > > suggested at this beginning of this thread did solve the link problem
> > > > and the resulting kernel ran for several hours. However, I think the
> > > > patch is still probably incorrect.
> > > >
> > > > I'm going to repost what I said in another message I sent today, this
> > > > time with a wider audience:
> > > >
> > > > The problem is that the name "sigprocmask" is getting renamed to
> > > > "kernel_sigprocmask" by a compiler directive in arch/um/Makefile, then
> > > > that name gets mangled into "sys_kernel_sigprocmask" by the
> > > > SYSCALL_DEFINE3(sigprocmask, ...) macro in kernel/signal.c.
> > > >
> > > > So, instead of the patch suggested earlier, I added the following line
> > > > to arch/um/sys-i386/sys_call_table.S:
> > > >
> > > > #define sys_sigprocmask sys_kernel_sigprocmask
> > > >
> > > > This made it compile and link correctly. Look at the symbols generated
> > > > by the compile of signal.c to see what I mean:
> > > >
> > > > # nm kernel/signal.o | grep sigprocmask
> > > > 0000008f r __kstrtab_kernel_sigprocmask
> > > > 00000040 r __ksymtab_kernel_sigprocmask
> > > > 00001ea6 T kernel_sigprocmask
> > > > 00002d67 T sys_kernel_sigprocmask
> > > > 00001faf T sys_rt_sigprocmask
> > > >
> > > > Unfortunately, it's a mystery to me that others haven't run into this
> > > > before. My host environment is RHEL 4 inside some kind of chroot.
> > >
> > > I've just started seeing this problem with some 2.6.29-rc3 kernel...
> > >
> > > Before, I did not have this problem with various 2.6.28-rc8 and 2.6.29-rc1
> > > kernels (and several older versions I don't remember).
> > >
> > > Given 2.6.29-rc1 works for me and 2.6.28.1 fails for you, I'm inclined to
> > > believe 2.6.28 is OK. I'll give it a try...
> > >
> > > BTW, I'm using CentOS 5.2.
> >
> > Following up from home...
> >
> > Indeed, 2.6.28 works, 2.6.28.1 doesn't.
> >
> > According to git bisect, it got introduced by the system call security fixes
> > (CVE-2009-0029), more specifically by this part:
> >
> > | commit fe7c0d987fb2cce464d29eec9dfcca6296b5eed7
> > | Author: Heiko Carstens <heiko.carstens@...>
> > | Date: Wed Jan 14 14:14:06 2009 +0100
> > |
> > | System call wrappers part 04
> > |
> > | commit b290ebe2c46d01b742b948ce03f09e8a3efb9a92 upstream.
> > |
> > | Signed-off-by: Heiko Carstens <heiko.carstens@...>
> > | Signed-off-by: Greg Kroah-Hartman <gregkh@...>
> > |
> > | --- a/kernel/signal.c
> > | +++ b/kernel/signal.c
> > | @@ -2425,8 +2424,8 @@ sys_sigpending(old_sigset_t __user *set)
> > | /* Some platforms have their own version with special arguments others
> > | support only sys_rt_sigprocmask. */
> > |
> > | -asmlinkage long
> > | -sys_sigprocmask(int how, old_sigset_t __user *set, old_sigset_t __user *oset)
> > ^^^^^^^^^^^^^^^
> > | +SYSCALL_DEFINE3(sigprocmask, int, how, old_sigset_t __user *, set,
> > ^^^^^^^^^^^
> > | + old_sigset_t __user *, oset)
> > | {
> > | int error;
> > | old_sigset_t old_set, new_set;
> >
> > Hence it allows sigprocmask to be redefined to kernel_sigprocmask by the C
> > preprocessor...
> >
> > This got backported to 2.6.27.12 as well, confusing people who ran post-2.6.27
> > development kernels and never noticed the problem (including Jeff and me)...
> >
> > It showed up in a "development" kernel in 2.6.29-rc2 only.
>
> Is there a real patch in Linus's tree for this fix that I can add to the
> .27 and .28 -stable kernel trees?
No, not yet. I just wanted to let you know about this regression.
I verified that Shane's solution:
#define sys_sigprocmask sys_kernel_sigprocmask
works for me, but that's definitely not the cleanest way.
Al Viro also had a suggestion to rework the SYSCALL_DEFINE* macros, but I
haven't tried it yet.
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@...
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
|